Preview: Dung-Thang-Hung - Argument Based Decision Making and Negotiation in E-business, Contracting a Land Lease for a Computer Assembly Plant

Attention! This is a preview.
Please click here if you would like to read this in our document viewer!


Argument-based Decision Making and Negotiation in E-business: Contracting a Land Lease for a Computer Assembly Plant Phan Minh Dung Phan Minh Thang Nguyen Duy Hung Department of Computer Science, Asian Institute of Technology GPO Box 4, Klong Luang, Pathumthani 12120, Thailand dung@cs.ait.ac.th, thangfm@ait.ac.th, nguyenduy.hung@ait.ac.th Abstract. We describe an extensive application of argument-based decision making and negotiation to a real-world scenario in which an investor agent and an estate manager agent negotiate to lease a land for a computer assembly factory. Agents are equipped with beliefs, goals, preferences, and argument-based decision-making mechanisms taking uncertainties into account. Goals are classified as either structural or contractual. The negotiation process is divided into two phases. In the first phase, following a recently proposed framework [8] the investor agent find suitable locations based on its structural goals such as requirements about

transportation; the estate manager agent determines favored tenants based on its structural goals such as requirements about resource conservation. In the second phase, we introduce a new novel argumentbased negotiation protocol for agents to agree on contract to fulfill their contractual goals such as waste disposal cost. 1 Introduction Argument-based negotiation enables agents to couple their offers with arguments, thus is believed to improve the quality of deals in such contexts as ebusiness, resource allocation [5]. We describe an extensive application of argumentbased decision making and negotiation to a real-world scenario in which an investor agent and an estate manager agent negotiate to lease a land for a computer assembly factory. Agents are equipped with beliefs, goals, preferences, and argument-based decision-making mechanisms taking uncertainties into account. Beliefs is structured as assumption-based argumentation framework. Goals are classified as structural if they

are about static properties of purchased items or services; like a structural goal of an investor for leasing a parcel of land could be that its location is near a sea port. Goals are classified as contractual if they are about features subject to negotiation leading to the agreement of a contract; like a contractual goal for above lease is that the rental cost is lower than $.9/m2 /month. Preferences are given by numerical rankings on goals. The negotiation process is divided into two phases. In the first phase, following a recently proposed contract negotiation framework [8] the investor agent finds suitable locations based on its structural goals; the estate manager agent determines favored tenants based on its structural goals. In the second phase, agents negotiate to agree upon a contract fulfilling their contractual goals. Agents starts negotiation about a basic item or a main service. As negotiation proceeds, agents may introduce sub-items or new services to accommodate each

other’s needs for a better deal. For example, the estate manager offers a waste disposal service at low price to make the land lease more attractive for the investor. This kind of reward is very common in daily business. To handle this pattern of negotiation, we develop a reward-based minimal concession negotiation protocol extending the original protocol [8], which does not deal with changes of negotiated items/services during negotiation. Like its predecessor, the new protocol ensures an efficient and stable agreement. The paper is structured as follows. Section 2.1 gives background on argumentbased decision making and section 2.2 presents our new negotiation protocol. Section 3 instantiates the contract negotiation framework [8] to model the decision making of an investor (we omit the estate manager’s part due to the lack of space). Section 4 is a design for implementation, and is followed by the conclusions. 2 2.1 Argument-based decision making and negotiation Argument-based

decision making An ABA framework,see [4, 7, 6, 8, 13] for details, is defined as a tuple hL, R, A, where i – (L, R) is a deductive system, consisting of a language L and a set R of inference rules, – A ⊆ L, referred to as the set of assumptions, is a (total) mapping from A into L, where x is referred to as the contrary – of x. We assume that the inference rules in R have the syntax l0 ← l1 , . . . ln (for n ≥ 0) where li ∈ L. Assumptions in A do not apprear in the heads of rules in R. A backward deduction of a conclusion x supported by a set of premises P is a sequence of sets S1 , ..., Sm , where S1 = {x}, Sm = P , and for every i, where y is the selected sentence in Si : If y is not in P then Si+1 = Si − {y} ∪ S for some inference rule of the form y ← S ∈ R. Otherwise Si+1 = Si . An argument is a (backward) deduction whose premises are all assumptions. In order to determine whether a conclusion (set of sentences) should be drawn, a set of assumptions needs to

Attention! This is a preview.
Please click here if you would like to read this in our document viewer!


be identified providing an “acceptable” support for the conclusion. Various notions of “acceptable” support can be formalised, using a notion of “attack” amongst sets of assumptions whereby X attacks Y iff for some y ∈ Y there is an argument in favour of y supported by (a subset of) X. A set of assumptions is deemed – admissible, iff it does not attack itself and it counter-attacks every set of assumptions attacking it; – preferred, iff it is maximally admissible. We will use the following terminology: – a preferred set of assumptions is called preferred extension. – a preferred extension of hL, R, A, i∪{a}, for some a ∈ A, is a preferred extension of hL, R, A, i containing a. – given a preferred extension E and some l ∈ L, E |= l stands for “there exists a backward deduction for l from some E ′ ⊆ E”. Agents are equipped with beliefs, goals, and preference. Following [8], an agent is defined as a tuple < G, B, P >, where – G ⊆ L is

its goal-base consisting of two disjoint subsets: G = Gstruct ∪ Gcontr , where Gstruct contains structural goals concerning the attributes of purchased items or services, for example a structural goal for leasing a parcel of land could be that its location is near a sea port; and Gcontr contains contractual goals concerning the contractual features of purchased items or services, for example a contractual goal for above lease is that the rental cost is lower than $.9/m2 /month. – P is its preference-base mapping goals from G to the set of natural number, ranking goals according to their importance so that the higher the number assigned to a goal, the more important the goal. – B is its belief-base represented by an ABA framework hL, R, A, i, where • R = Ri ∪ Rn ∪ Rc , where ∗ Ri represents information about concrete items or services to be traded, for example the distance from a parcel of land to a sea port is 30 kms. ∗ Rn consists of rules representing (defeasible)

rules or norms, for example textile industries require only low skilled labour force. ∗ Rc represents information related to contractual goals, for example an estate manager often offers rental discount for investors in electronics. • A = Ad ∪ Ac ∪ Au , where ∗ Ad consists of assumptions representing items or services for transactions, for example location1, location2. ∗ Ac represents control assumptions related to defeasible norms. ∗ Au contains assumptions representing the uncertainties about items or services to be traded, for example whether the labour skill available at a location is high. A contract is viewed as a transaction between agents playing different roles, characterized by an item or service package and an assignment of values to item attributes. Formally, a contract between two agents is a tuple < Buyer, Seller, Item, F eatures > where – Buyer, Seller are different agents representing the buyer and seller in the contract – Item is the item or

service package to be traded in the transaction – F eatures is an assignment of values to item/service attributes An example contract is < investor, estate manager, location2, rental = $1.0/m2/month > indicating that the estate eatate manager leases location2 to the investor at $1.0/m2/month. To agree on a contract, agents engage in a two-phase negotiation process. In the first phase, the buyer agent evaluates available items or services to determine how they satisfy its needs. In the second phase, the buyer agent negotiates with the seller for items/services that have passed the first phase. Choices in the first phase are available items or services, and choices in second phase are possible deals. The value of a choice is represented by the set of goals satisfied by the choice. Let d ∈ Ad be a choice available to an agent < G, B, P > and g ∈ Gstruct , we says that – g is credulously satisfied by d if there is a preferred extension E of B ∪ {d} such that E |= g

– g is skeptically satisfied by d if, for each preferred extension E of B ∪ {d}, E |= g The framework in [8] models risk-averse decision makers who consider the value of choice d, denoted by V al(d) as the set of goals skeptically satisfied by d. Definition 1. Let d, d′ be two choices and s = V al(d),s′ = V al(d′ ) be the sets of goals representing the values of d, d′ respectively. Then d is preferred to d′ , denoted by d ⊒ d′ iff – there exists a goal g that is satisfied in s but not in s′ , and – for each goal g ′ , if P (g ′ ) ≥ P (g) and g ′ is satisfied in s′ than g ′ is also satisfied in s. That is, choices enforcing higher-ranked goals are preferred to those enforcing lower-ranked goals. 2.2 A Reward-based Minimal Concession Negotiation Protocol Suppose an investor and an estate manager consider a partnership. The estate manager wants to provide not only land lease but also other estate services such as wastewater treatment. However, at the

Attention! This is a preview.
Please click here if you would like to read this in our document viewer!


beginning the estate manager may not have full information about the investor’s needs, and the investor may also not have full information about the estate services. They often start negotiation about only land lease. As negotiation proceeds, the estate manager may introduce additional services when he discovers the investor’s needs. These services are called value-added if their values are lower than that of the main service, however they are offered at significantly lower prices than the prices that could be obtained if purchased separately from different service providers. This kind of reward is very common in business when a service provider offers an extra service at low price to increase the attractiveness of a service package. For example in our scenario the estate manager can offer waste disposal service for some kinds of industrial waste produced from the manufacturing of printed circuits when he discovers this investor’s need. The reason the estate manager offers

this service at low price is that he collects similar wastes from other tenants as well and then treats them in large scale. Furthermore if the investor contracts with an outside company, he has to pay extra cost for transportation. To handle this pattern of negotiation, we extend the minimal concession protocol introduced in [8], which is itself inspired by the monotonic concession protocol in [23]. We assume that a buyer agent β needs a main service msr and a set Sβ of (value-added) services. After the first phase, the buyer agent decides to start negotiation to buy msr from a seller σ and buys other services in Sβ from wherever the best offers he gets. The seller σ wants to sell the main service msr, possibly packaged with other services in Sσ possibly different from Sβ . A service package is defined as a set p = {msr} ∪ r, where r ⊆ Sσ ∩ Sβ . Agents negotiate to determine a concrete service package for transaction. The value of such transaction is defined by a

contractual state. Definition 2. A contractual state is a pair hp, assi where p is a service package and ass is an assignment of values to contractual attributes(e.g. {price = 10K, deliveringT ime = 1week}). The set of all contractual states is denoted by CS while the set of all contractual states about p is denoted by CSp . The preference of an agent α between contractual states can be represented as a total pre-order ⊒α where t ⊒α t′ states that t is preferred to t′ (for α). ⊒α is assumed to be consistent with the partial order obtained from Definition 1. We assume that agent knows its preferences between contractual states. Agents are not assumed to know the preferences between contractual states of other agents except if the states have the same package. We say that: t is strictly preferred to t′ for agent α, denoted by t ⊐α t′ if t ⊒α t′ and t′ 6⊒α t; t is equally preferred to t′ for agent α, denoted by t =α t′ if t ⊒α t′ and t′ ⊒α

t; t dominates t′ , denoted by t > t′ if t is preferred to t′ for both seller and buyer (i.e. t ⊒β t′ and t ⊒σ t′ ) and, for at least one of them, t is strictly preferred to t′ ; t is Pareto-optimal if it is not dominated by any other contractual state. We also assume that each agent α possesses an evaluation function λα that assigns to each package p a contractual state λα (p) representing the reservation value of p for α. For the buyer agent β (or the seller σ, resp.), λβ (p) (or λσ (p), resp.) is the maximal (or minimal, resp.) offer it could make (or accept, resp.). The possible deals (contracts) that agent α could accept for a package p is defined by P Dα (p) = {t|t ∈ CSp and t ⊒α λα (p)}. Furthermore, agents are rational in the sense that they would not accept a deal that is not Pareto-optimal. We define the negotiation set N S(p) (about a package p) as the set of all Paretooptional contractual states in P Dβ (p) ∩ P Dσ (p). It is

not difficult to see that for t, t′ ∈ N Sp , t′ ⊐σ t iff t ⊐β t′ , t′ ⊒σ t iff t ⊒β t′ , and t′ =σ t iff t′ =β t. A package p is said to be negotiable if N S(p) is not empty. It follows that p is negotiable iff λβ (p) ⊒σ λσ (p) (or λσ (p) ⊒β λβ (p)). We represent a state of a negotiation as a tuple h(σ, vσ ), (β, vβ )i where vσ , vβ are the lastest offers of the seller agent and the buyer agent respectively. Offers are represented by contractual states. Agent starts negotiation by putting forwards its most preferred offer from the initial negotiation set N S({msr}). That is, the seller agent offers to sell msr at λβ ({msr}) and the buyer agent offers to buy it at λσ ({msr}). The negotiation state after these moves is h(σ, h{msr}, λβ ({msr})i), (β, h{msr}, λσ ({msr})i)i. Suppose now that agents are negotiating about a package p and the current negotiation state is h(σ, vσ ), (β, vβ )i. If agent α ∈ {β, σ} taking its

Attention! This is a preview.
Please click here if you would like to read this in our document viewer!


turn to move next puts an offer v about the package p then v should be an element of N S(p) such that v ⊒α vα because when an agent makes a new offer, it should be at least as preferred for its opponents (denoted by α) as the one it has made previously. Instead of making new offer for package p, the agent could introduce or request a set of new services r to be included in the negotiation. To determine the asking price for the new package from the current stage of negotiation, we α assume that each agent α possesses a function fp,r : CSp → CSp∪r computing its first offer for p ∪ r from an offer about p, where r ∩ p = ∅. It is sensible to α assume that fp,r satisfies following constraints 1. Lossless. The new offer is strictly preferred (or preferred, resp.) to its previous offer for agent α (or α, resp.), who introduces/requests (or who replies, resp.) α α fp,r (v) ⊐α v and fp,r (v) ⊒α v 2. Reward. The seller offers r additional to p at price cheaper than

the price the buyer could get r from other vendors. σ fp,r (v) ⊐β fp,r (v) ⊒β v, where fp,r is a function returning the minimal possible cost of p ∪ r if the buyer purchases p at v from the seller and then purchases r from other vendors. 3. Monotonicity. Service inclusion retains preference order. α α α α If v2 ⊐α v1 then fp,r (v2 ) ⊐α fp,r (v1 ). If v2 =α v1 then fp,r (v2 ) =α fp,r (v1 ). 4. Value-added. Service inclusion expands negotiation space. α α fp,r (v) ⊐α fp,r (v). 5. Flatten. The inclusion of a set r of services can be substituted by the inclusions of individual services in r consecutively. α α α α ∀asr ∈ r, fp,r (v) = fp∪{asr},r−{asr} (fp,{asr} (v)); and fp,∅ (v) = v Example 1. To motivate and explain the above constraints, let’s consider a simple case where cost (say in US$) is the only contractual attribute. A contractual state is defined by a pair hp, vi where v a natural number representing a cost of package p. So λα (p) = hp,

Γα (p)i, where Γα (p) is a natural number representing the reservation cost of p for α. It is reasonable to assume that Γβ (p ∪ r) = Γβ (p) + Γβ (r) since β may have to buy each services separately from different vendors. Suppose the lowest price the seller is willing to offer r is d(r), which could be considered as a fixed effective reservation price of r. Hence Γσ (p ∪ r) = Γσ (p) + d(r). So it is sensible for the buyer to set β σ fp,r (hp, ni) = hp∪r, n+d(r)i; and for the seller to set fp,r (hp, ni) = hp∪r, n+Γ (r)i where Γ (r) is smaller than Γ0 (r) which is the minimal possible amount the buyer has to pay for getting r on the market from other vendors (i.e. the minimal possible sum of market price of r and cost for packaging p, r together). It is sensible to expect Γβ (r) ≥ Γ0 (r) (since β has to pay at least the minimal market price of r and packaging cost in order to obtain r from other vendors), and Γ (r) > d(r) and fp,r (hp, ni) = hp

∪ r, n + Γ0 (r)i. It is easy to see the satisfaction of above constraints when written in simplified forms belows β σ – Lossless: fp,r (hp, ni) = hp ∪ r, n + d(r)i ⊐β hp, ni and fp,r (hp, ni) = hp ∪ r, n + Γ (r)i ⊐σ hp, ni. σ – Reward: fp,r (hp, ni) = hp ∪ r, n + Γ (r)i ⊐β fp,r (hp, ni) = hp ∪ r, n + Γ0 (r)i ⊒β hp, ni – Monotonicity: If n1 < n2 then hp ∪ r, n1 + d(r)i ⊐β hp ∪ r, n2 + d(r)i and hp ∪ r, n2 + Γ (r)i ⊐σ hp ∪ r, n1 + Γ (r)i – Value-added: hp P ∪ r, n + d(r)i ⊐β hp ∪ r, n + ΓP (r)i – Flatten: d(r) = asr∈r d({asr}) and Γ (r) = asr∈r Γ ({asr}) After the inclusion of new services r to the current package p, the current σ β negotiation state is changed to h(σ, fp,r (vσ )), (β, fp,r (vβ ))i. From the above constraints, it follows that σ – The negotiation space is changed from {v|vσ ⊒σ v ⊒σ vβ } to {v|fp,r (vσ ) ⊒σ β σ σ β v ⊒σ fp,r (vβ )}, which is not empty since fp,r (vσ )

⊐σ fp,r (vβ ) ⊐σ fp,r (vβ ). σ β β σ – ∀v ∈ N S(p), fp,r (v) ⊐σ fp,r (v) ⊒σ v and fp,r (v) ⊐β fp,r (v) ⊐β v. Thus if agents could reach a deal about p then they could reach a new deal about p ∪ r that dominates the other. σ – the size of {t|fp,r (v) ⊒β t ⊐β fp,r (v)} (or Γ0 (r) − Γ (r) as in example 1) could be considered as part of a reward from the seller to the buyer. We define reward-based monotonic concession negotiation as an interleaving sequence of concession negotiation about the package already accepted for negotiation and negotiation for service inclusion. Concession negotiation is an alternating sequence of moves between the seller agent and the buyer agent. Suppose that agents are negotiating about a current package p with the negotiation state h(σ, vσ ), (β, vβ )i. A move is represented by a tuple htype, α, vi, where type is type of the move, α is the agent making the move, and v is an element of the current negotiation

Attention! This is a preview.
Please click here if you would like to read this in our document viewer!


space N S(p). If v is strictly preferred to the agent’s previous offer for its opponent (v ⊐α vα ), then type is concede; otherwise, it is standstill. After a buyer’s (or seller’s, resp.) concession move hconcede, β, vi (or hconcede, σ, vi, resp.) the current negotiation state is changed to h(σ, vσ ), (β, v)i (or h(σ, v), (β, vβ )i, resp.). Negotiation about the inclusion of a set r of new services can be initiated with an introduction move for r of the seller agent or a request move for r of the buyer agent. An introduction move is represented σ by a tuple hintroduce, σ, fp,r (vσ )i. A request move is represented by a tuple hrequest, β, hp ∪ r, ⊥ii where ⊥ means that the buyer is asking the seller to state its price. The buyer (or seller, resp.) will reply to the introduction (or request, resp.) move by making a reply move. If the buyer agent needs a subset r′ = r∩Sβ of introduced services, it will reply positively by making a positive reply move,

β represented by a tuple hreply, β, fp,r ′ (vβ )i. Similarly, if the seller agent provides ′ a subset r = r ∩ Sσ of requested services, it will make a positive reply move, σ represented by a tuple hreply, σ, fp,r ′ (vσ )i. A positive reply move will change β σ the current negotiation state to h(σ, fp,r ′ (vσ )), (β, fp,r ′ (vβ ))i. Agents could reply negatively by repeating its last offer to indicate that the proposal for service inclusion fails and the negotiation state remains unchanged. Formally, a reward-based monotonic concession negotiation is a sequence m1 , m2 , . . . , mn of alternative moves of the form mi =< typei , αi , vi > between a buyer agent and a seller agent where the seller agent starts the negotiation by offering to sell the main package at the buyer’s reservation value, and the buyer agent replies by offering to buy it at the seller’s reservation value. Suppose now that the current negotiation state is h(σ, vσ ), (β, vβ )i.

Subsequent moves mn , n ≥ 3 could be of one of the types introduction, request, reply, standstill, or concession, where – If mn is an introduction move of the seller agent (or a request move of the σ buyer agent, resp.) for a set r of new services, then mn = hintroduce, σ, fp,r (vσ )i (or mn = hrequest, β, hp ∪ r, ⊥ii, resp.) where p ∩ r = ∅. The current state of negotiation remains unchanged. – If mn is a positive reply move of the seller (or buyer, resp.) agent then the previous move is a request move of the buyer agent (or introduction move of the σ seller agent, resp.) for a set r of new services and mn = hreply, σ, fp,r ′ (vσ )i β ′ ′ (or mn = hreply, β, fp,r′ (vβ )i, resp.) where r = r ∩ Sσ (or r = r ∩ Sβ , reps.) β σ and r′ 6= ∅. The new negotiation state is h(σ, fp,r ′ (vσ )), (β, fp,r ′ (vβ ))i. – If mn is a negative reply move of the seller (or buyer, resp.) agent then the previous move is a request move of the buyer (or

an introduction move of the seller, resp.) agent for a set r of new services and r ∩ Sσ = ∅ (or r ∩ Sβ = ∅, resp.) and mn , mn−2 coincide with exception of their types. The current negotiation state remains unchanged. – If mn is a standstill move then the previous move mn−1 is not an introduction/request move and mn = hstandstill, αn , vα i – If mn is a concession move then mn = hconcede, αn , vn i and the previous move mn−1 is not an introduce/request move, and • if αn is the seller agent then vn ⊐β vσ and the new negotiation state is h(σ, vn ), (β, vβ )i • if αn is the buyer agent then vn ⊐σ vβ and the negotiation state is h(σ, vσ ), (β, vn )i – A service should not be requested or introduced twice. A seller’s positive reply move or introduction move for a set r of new services where r ∩ Sβ 6= ∅ is basically an argument about a reward for the buyer agent represented in a short form. However, the move is not seen as a seller’s

concession since it does not suffer any loss in comparison with its previous offer. A negotiation terminates successfully if one of the agents accepts an offer. The seller (or buyer, resp.) agent accepts an offer made in a concession move mn = hconcede, αn , vn i by the buyer agent (or seller agent, resp.) if vn ⊒σ vσ (or if vn ⊒β vβ , resp.) where h(σ, vσ ), (β, vβ )i is the negotiation state after mn−1 . A negotiation terminates with failure if the agents make three standstill moves consecutively. Two standstills are said to be consecutive if moves between them are only introduction, request, and negative reply moves. Definition 3. If a concession move leads to a successful termination, then the move is called a finishing move. Definition 4. A contractual state t′ is said to be a minimal concession of agent α wrt t about a package p if t, t′ ∈ N S(p) and t′ is strictly preferred to t for α and for each contractual state r ∈ N S(p), if r is strictly

Attention! This is a preview.
Please click here if you would like to read this in our document viewer!


preferred to t for α then t′ is preferred to r for α. Definition 5. A contractual state t′ is said to be a hasty concession of agent α wrt t about a package p if t′ is a minimal concession of α wrt t about p and α there exists a service asr ∈ / p such that a minimal concession of α wrt fp,{asr} (t) α ′ (about p ∪ {asr}) is strictly preferred to fp,{asr} (t ) for α. So if agent α makes a minimal concession move following an introduce/request move for asr, then it will reach a state which is preferred for it to the state if it makes the introduce/request move for asr following a minimal but hasty concession move. Definition 6. Reward-based minimal concession negotiation is a reward-based monotonic concession negotiation where each agent only makes a minimal concession in a concession move and no agent makes a hasty concession move or a finishing move if it can make a request or introduction move. Furthermore agent standstills only if its opponent standstills in

previous step. The following proposition shows that request and introduction moves represent a simple but effective information-seeking dialogs (for honest agents). Proposition 1. If both seller agent σ and buyer agent β negotiate using the reward-based minimal concession strategy, then negotiation terminates successfully by a deal containing all services in Sσ ∩ Sβ . Let deal(stσ , stβ ) denote the deal of negotiation if σ, β use reward-based monotonic concession strategies stσ , stβ respectively. If the negotiation terminates in failure then deal(stσ , stβ ) is assigned a special value ⊥, which is less preferred to any deal for both agents. Proposition 2. For any reward-based minimal concession strategy stσ (or stβ , resp.) and any reward-based monotonic concession strategy stβ (or stσ , resp.), there exists a reward-based minimal concession strategy st′β (or st′σ , resp.)such that deal(stσ , st′β ) ⊒β deal(stσ , stβ ) (or deal(st′σ , stβ )

⊒σ deal(stσ , stβ ),resp.). The following proposition shows that reward-based minimal concession strategies are equivalent. Proposition 3. If stσ , stβ , st′σ , st′β are reward-based minimal concession strategies then deal(stσ , stβ ) = deal(st′σ , st′β ) A strategy is said to be in symmetric Nash equilibrium [14] if under the assumption that if one agent uses this strategy the other agent can not do better by not using this strategy. A strategy is said to be in symmetric subgame perfect equilibrium [15] if for each history of negotiation h, under the assumption that one agent uses this strategy starting from h, the other agent can not do better by not using this strategy starting from h. It is not difficult to see: Theorem 1. The reward-based minimal concession strategy is in symmetric Nash equilibrium and symmetric subgame perfect equilibrium. 3 Investor’s decision making Foreign investors often lease serviced land plots inside industrial estate to set

up factories [1, 20, 18]. In this session we examine how an investor in computer assembly selects a location from Vietnam industrial property market. 3.1 The investor Suppose an investor has analyzed computer market demand and decided to invest in assembly of low-end computers. To set up a computer assembly plant, the investor has to make decisions about technologies to be used in the plant, and location of the plant. Goals of the investor. The investor wants to achieve – structural goals related to technology choices, for example • (g1 ) capacity of the plant could be easily adjusted to adapt to market demand • (g2 ) enhancing the dynamics of assembly line (see norm2 below) – structural goals related to the location of the plant, for example • (g3 ) qualified labour force is available at the location • (g4 ) average wage does not exceed some threshold, e.g. $1.3/hour • (g5 ) the location is near a sea port • (g6 ) the location is eligible for sufficient government

investment incentives – contractual goals related to industrial estate services • (g7 ) reservation price for land lease is $.9m2 /month • (g8 ) reservation price for waste disposal is $.3m2 /month The investor determines the preferences over goals by ordering them according to their importance, for example, g3 ⊒ g1 ⊒ g6 ⊒ g2 , g5 , g9 , h3 ⊒ g4 , and encodes the order by numerical rankings such as P (g1 ) = 5, P (g2 ) = 3, P (g3 ) = 6, P (g4 ) = 1, P (g5 ) = 3, P (g6 ) = 4, P (g7 ) = 3, P (g8 ) = 3. High ranked goals include labour and capacity adjustment. This is because computer assembly mainly concerns manual operation, so labour takes important role. Capacity needs to be adjusted according to very high expected demand variability. Lower ranked goals include wage and sea port. This is because average wage in Vietnam is very low and transportation cost is not big in comparison with the computer price. Knowledge about technology choices. Knowledge about technologies

Attention! This is a preview.
Please click here if you would like to read this in our document viewer!


demonstrates the technical know-how of the investor. The most important decision about technology in computer assembly concerns the structure of the assembly process. There are two kinds of assembly lines[10]. In parallel line, the whole assembly process is completed by a small group of workers at one workstation. In serial line, the assembly process is divided into sub-processes which are completed at different workstations in a specific order. The investor should know the influence of a technology choice on his goals. For example, to decide between parallel or serial lines, he should be aware of the relations between different factors: – norm1 related to g1 (capacity adjustment). If market demand changes rapidly, the investor needs to be able to adjust production capacity quickly. In parallel line, increasing capacity requires a duplication of workstations. In serial line, increasing capacity requires adding more workers to assembly line. Hence, capacity adjustment in parallel line

incurs the cost of redundant workstations while in serial line it incurs the cost of modifying working procedures – norm2 related to g2 (line dynamics). In serial line, workers in a workstation work under pressure from workers in other workstations of the same line. Workers at a workstation may work faster if they know that the next workstation is idle or they have just taken longer time completing the last unit, for fear that they are holding up the line. The effect of this behavior is that the line speed is maintained by workstations pushing and pulling material through the line, possibly enabling higher throughput than parallel line where no such inter-workstation pressure exists. This advantage of serial line is referred to as the line dynamics. – norm3 related to g3 (labour availability). Parallel line requires higher labour force skill than serial line. In parallel line, a worker is responsible for the assembly of the whole unit while in serial line, he is just responsible

for completing tasks assigned to his workstation. • (norm3.a ) The investor classifies labour force skill of a location into low or high. Parallel line requires high skill labour force while serial line just requires low skill labour force 1 1 We assume a two-level classification for simplicity. The classification could be more than two. • (norm3.b ) The investor could improve labour force skill by organizing training programs in electronics. – rule1 (uncertainties about labour skill). If there is no information about labour skill of a location, the investor can assume that it could be either high or low. The investor should know factual information about technology, for example – f act1 related to norm1 • (f act1.a ) computer assembly only requires manual tools and inexpensive general purpose workstations. The cost for factory floor for redundant workstations is not significant. Thus the cost of redundant workstations for a capacity buffer can be ignored. • (f act1.b

) changing working procedure when workers are added or removed from serial line incurs a significant throughput loss because the line takes sometime to stabilize. It follows from norm1 and f act1 that the investor can easily adjust the capacity of the plant if he selects parallel line. However, it is costly to do so if he selects serial line. Knowledge about locations. The investor develops a set of criteria to evaluate suitability of a location as follows – norm4 related to g3 (labour availability). Labour availability of a location is assessed by its population, e.g. greater than 40K, and its labour force qualification. Labour force is qualified when its labour force skills meets the requirements posed by selected technology. – norm5 related to g5 (sea port accessibility). Location should be connected to a sea port by national roads with distance smaller than, e.g. 35km to reduce transportation cost because some computer components need to be imported by sea. – norm6 related to

g6 (incentives). Tax reduction for at least five years is considered as an attractive incentive. Information of land plots for lease could be as follows – location1: population is 45K; distance to sea port is 30km by national road; average wage is $1/hour; tax reduction in the first three years of any investment; there is an on-site electronics training center. – location2: population is 46K; distance to sea port is 35km by national road; average wage is $1.5/hour; tax reduction in the first eight years of any investment; the estate manager is considering building either a training center in electronics or a mansion on site; IT industry is encouraged by rental cost reduction. – rule2 (uncertainties about estate facilities). If there is no information about whether the estate manager (of location2) is going to develop a mansion or a training center, the investor assumes that he could develop either facility. Decision analysis.With two candidate locations and a technology

Attention! This is a preview.
Please click here if you would like to read this in our document viewer!


choice between serial and parallel, the investor has four options as follows: (location1, serial), (location1, parallel), (location2, serial), (location2, parallel). The satisfaction of goals wrt the above four options are summarized in Table 1. Option (location1, parallel) satisfies goal g1 because of norm1 and f act1 ; and dissatisfies g2 because of norm2 . Option (location2, parallel) neither satisfies nor dissatisfied goal g3 . This is because there is no information about labour skill, nor a planing to build training center or mansion of location2. By rule2 , the investor infers that estate management could build either facility. By rule1 , the investor can assume that the labour skill is either high or low. If the investor believes the labour skill is high, or a training center will be built, then by norm3 and norm4 , goal g3 will be satisfied; otherwise, g3 will be dissatisfied. Thus, for a risk-averse agent, option Table 1. Investor’s goal satisfaction (location1, parallel)

is the most favoured. 3.2 Formal representation of the investor agent The investor agent is represented by a triple < G, P, B > where, struct contr – goal-base G = Gstruct , where location ∪ Gtech ∪ G struct • Gtech consists of structural goals related to technology choices: (g1 ) capacityAdjustment; (g2 ) lineDynamics. • Gstruct location consists of structural goals related to location: (g3 ) labourAvailability; (g4 ) wage < $1.3/hour; (g5 ) seaP ortAccessibility; (g6 ) incentives. • Gcontr consists of contractual goals: (g7 ) rental ≤ $.9/m2/month; (g8 ) wasteDisposal ≤ $.3/m2 /month. – preference-base : P (g1 ) = 5, P (g2 ) = 3, P (g3 ) = 6, P (g4 ) = 1, P (g5 ) = 3, P (g6 ) = 4, P (g7 ) = 3, P (g8 ) = 3. – the belief-base B is an ABA framework hL, R, A, i, where • R = Ri ∪ Rn ∪ Rc ∪ Rf ,where ∗ Ri represents information about locations Representation of information about location 1: pop = 45K ← location1; distanceT oSeaP ort = 30km ←

location1; nationalRoad ← location1; wage = $1/hour ← location1; yearsOf T axReduction(3) ← location1; trainingCenter ← location1. Representation of information about location 2: pop = 46K ← location2; distanceT oSeaP ort = 35km ← location2; nationalRoad ← location2; wage = $1.5/hour ← location2; yearsOf T axReduction(8) ← location2. ∗ Rn = Rntech ∪ Rnlocation , where · Rntech consists of representation of norms about technologies as well as rules representing uncertainties Representation of norm1 : capacityAdjustment ← parallel, asm1; dif f icultT oAddW orkstations ← highCostF orRedundantW orkstations; capacityAdjustment ← serial, asm2; dif f icultT oAddW orkers ← expensiveT oChangeP rocedure. Representation of (the conclusion in) norm2 : lineDynamics ← serial. Representation of norm3 : qualif ication ← parallel, highSkill; qualif ication ← serial; highSkill ← trainingCenter. Representation of rule1 : lowSkill ← notHighSkill; highSkill ←

notLowSkill. Representation of rule2 : mansion ← notT rainingCenter, location2; trainingCenter ← notM ansion, location2. · Rnlocation consists of representation of norms about location Representation of norm4 : labourAvailability ← pop > 40K, qualif ication. Representation of norm5 : seaP ortAccessibility ← distanceT oseaP ort < 35km, nationalRoad. Representation of norm6 : incentives ← yearsOf T axReduction(X), X ≥ 5. ∗ Rf consists of Representation of f act1 : lowCostF orRedundantW orkstations ←; expensiveT oChangeP rocedure ←. • A = Ad ∪ Ac ∪ Au , where ∗ Ad = Atech ∪ Alocation , where d d tech · Ad = {serial, parallel} are assumptions representing technology choices serial = parallel; parallel = serial. · Alocation = {location1, location2} are assumptions representing d location choices location1 = location2; location2 = location1. ∗ Ac = {asm1, asm2} are control assumptions related to norms asm1 = dif f icultT oAddW orkstations; asm2 = dif f

icultT oAddW orkers. ∗ Au = {notLowSkill, notHighSkill, notT rainingCenter, notM ansion} are assumptions representing uncertainties about locations. notHighSkill = highSkill; notLowSkill = lowSkill; notT rainingCenter = trainingCenter; notM ansion = mansion. The investor agent’s decision analysis. Table 2 shows structural goal states and their min satisfied by the composite decision. For example, g2 is credulously satisfied by option (location1, serial) -assumptions contained in the preferred extension {location1, serial, asm1, notLowSkill}. As a risk-averse decision maker, the value of an option is the min of all its goal states. So, he considers the value of option (location2, parallel) is {g1 , g5 , g6 }. Table 2. Investor’s preferred extensions 4 Design for implementation An agent could be implemented by two separate modules. The first module is for internal decision making and the second is for bargaining. The first module is the direct translation of the agent formal

Attention! This is a preview.
Please click here if you would like to read this in our document viewer!


representation into CaSAPI 2 and MARGO 3 . The second module is the implementation of the reward-based minimal concession strategy. A sample fragment of the seller agent’s second module is as follows. We assume the agent possesses a function fconcede to compute its next minimally conceded offer, function noOfStandstills() to return the number of consecutive standstills in the negotiation, and function notHasty(Sσ , p, vσ , fconcede) defined σ σ as ∀asr ∈ Sσ .fp,{asr} (fconcede (vσ )) =σ fconcede(fp,{asr} (vσ )) to check if a concession is not hasty. 2 3 www.doc.ic.ac.uk/ dg00/casapi.html http://margo.sourceforge.net/ 1. The seller opens the negotiation by offering to sell the main service at the buyer’s reservation value. O(1,start,β, {msr}, λσ (s)) ← 2. The buyer replies by offering to buy it at the seller’s reservation value. O(2,start,σ, {msr}, λβ (s)) ← 3. Suppose now that the seller has its turn at the nth move in the negotiation, Sσ contains only

value-added services not introduced yet, and the negotiation after state mn−1 is hp, (σ, vσ ), (β, vβ )i. (a) If the buyer standstills then the seller also standstills. O(n,standstill,σ, p, vσ ) ← O(n − 1,standstill,β, , ) (b) If the buyer requests a set r of services then the seller replies i. positively if he can provide O(n,reply,σ, p ∪ r′ , V ) ← O(n − 1,request,β, p ∪ r, ), r′ = r ∩ Sσ , r′ 6= σ {}, V = fp,r ′ (vσ ) ii. negatively otherwise. O(n,reply,σ, p, vσ ) ← O(n − 1,request,β, p ∪ r, ), r ∩ Sσ = {} (c) If the buyer replies or concedes then i. if the seller has services then A. he either introduces O(n,introduce,σ, p∪r, V ) ← O(n−1, t, β, p, vβ ), t ∈ {reply,concede}, r ⊆ σ Sσ , V = fp,r (vσ ) B. or concedes provided that this is not a finishing or hasty concession move O(n,concede,σ, p, V ) ← O(n−1, t, β, p, vβ ), t ∈ {reply,concede}, V = fconcede(vσ ), vβ ⊐β V,notHasty(Sσ , p, vσ , fconcede) ii.

else, the seller concedes. O(n,concede,σ, p, V ) ← O(n−1, t, β, p, vβ ), t ∈ {reply,concede}, Sσ = {}, V = fconcede(vσ ) (d) The seller accepts an offer made in a concession move mn−1 of the buyer if vn−1 ⊒σ vσ StopAndAccept ← O(n − 1,concede,β, p, vn−1 ), vn−1 ⊒σ vσ (e) The negotiation terminates in failure if there are three consecutive standstills. StopInFailure ←noOfStandstills() = 3 The design of the module for bargaining of the buyer is similar. 5 Conclusion We have extended the two-phase contract negotiation framework[8] where by in the first phase a buyer agent decides on items fulfilling its structural goals, and in the second phase it negotiates with the agent selling the item determined in the first phase to agree on a contract. The new framework improves on its predecessor by allowing agents to exchange information about each other’s needs and capabilities during negotiation to change negotiated items. It also drops the assumption

that the seller has no structural goals (we do not present the seller’s part due to the lack of space). Our new framework, like its predecessor, allows agents to achieve Nash and subgame perfect equilibria. The first phase is supported by a decision-making mechanism using argumentation and preferences. A number of such decision-making mechanisms exist, e.g. [11, 16, 12, 3]. This argument-based framework can deal with decision making, uncertainties and negotiation. However, we have restricted ourself only to a simple and ideal case where we assume that the agents are honest and open to each other, and ignore the need of information-seeking in the first phase. The second phase is also supported by argumentation with only reward-based arguments. We plan to explore other types of arguments and define a communication machinery to support information-seeking in the future. We have illustrated our approach using a scenario studied in the ARGUGRID project 4 . We believe that our approach

could be fruitfully applied to scenarios where a buyer negotiates for a current item and plans possible subsequent encounters with the same seller for additional items. For example, negotiation between a car seller and buyer may cover possible after-sale services. Several works exist on argumentation-based negotiation [17]. For example, [21] propose a protocol and a communication language for dealing with refusals in negotiation. It would be useful to see how this protocol and communication language may be used to support the two-phase negotiation framework we have defined. Also, [2] presents an abstract negotiation framework whereby agents use abstract argumentation internally and with each other. Our framework instead is tailored to the very common case in business and assumes a very concrete and structured underlying argumentation framework. Our reward-based monotonic minimal concession strategy for fair agents is inspired by the monotonic concession protocol of [23], though it

Attention! This is a preview.
Please click here if you would like to read this in our document viewer!


differs from it in significant ways. In our framework the agent moves alternatively where in [23] they move simultaneously. The condition for terminating the negotiation is also different. As a result, the minimal concession strategy is in symmetric subgame perfect equilibrium in our framework while the corresponding strategy in [23] is not even in symmetric Nash equilibrium. We do not use an explicit function of utilities to calculate a notion of risk to determine the player who should make the next concession as in [23] as research into practical negotiation behavior [19] shows that a strategy of making a large concession and then expects the other player to match is not a sound practical negotiation strategy as the other player often then discounts such concession as not important to the player who made it. In this paper we have considered just two agents, not within multi-agent systems as in other existing works, e.g. [9, 22] and focused instead on the full negotiation process,

from the identification of issues to bargain about to the actual bargaining, thus linking argumentation-based decision making to the monotonic concession protocol. 4 www.argugrid.eu 6 Acknowledgements We thank the referees for constructive comments and criticisms. This work was partially funded by the Sixth Framework IST program of the European Commission under the 035200 ARGUGRID project. References 1. Amata. Amata Vietnam. www.amata.com, June 2008. 2. L. Amgoud, Y. Dimopolous, and P. Moraitis. A unified and general framework for argumentation-based negotiation. In Proc. AAMAS’2007, 2007. 3. K. Atkinson and T. Bench-Capon. Practical reasoning as presumptive argumentation using action based alternating transition systems. Artificial Intelligence, 171(10–15):855–874, 2007. 4. A. Bondarenko, P.M. Dung, R.A. Kowalski, and F. Toni. An abstract, argumentation-theoretic approach to default reasoning. Artificial Intelligence, 93(1-2):63–101, 1997. 5. Yann Chevaleyre, Paul E.

Dunne, Ulle Endriss, Jrme Lang, Michel Lematre, Nicolas Maudet, Julian Padget, Steve Phelps, Juan A. Rodrguez-aguilar, and Paulo Sousa. Issues in multiagent resource allocation. Informatica, 30:2006, 2006. 6. P.M. Dung. On the acceptability of arguments and its fundamental role in nonmonotonic reasoning, logic programming and n-person games. Artificial Intelligence, 77:321–357, 1995. 7. P.M. Dung, R.A. Kowalski, and F. Toni. Dialectic proof procedures for assumptionbased, admissible argumentation. Artificial Intelligence, 170:114–159, 2006. 8. P.M. Dung, P.M. Thang, and F. Toni. Towards argumentation-based contract negotiation. COMMA’08, 2008. 9. U. Endriss. Monotonic concession protocols for multilateral negotiation. In P. Stone and G. Weiss, editors, Proceedings of the 5th International Joint Conference on Autonomous Agents and Multiagent Systems (AAMAS-2006), pages 392–399. ACM Press, May 2006. 10. T.M. Furey. Decision elements in the design of a consumer electronics

assembly plant. Master’s thesis, Sloan School Of Management, Massachusetts Institute of Technology, May 1999. 11. A.C. Kakas and P. Moraitis. Argumentation based decision making for autonomous agents. In Proc. AAMAS’03, pages 883–890, 2003. 12. M. Morge and P. Mancarella. The hedgehog and the fox: An argumentation-based decision support system. In Proc. ArgMAS, 2007. 13. M. Morge and P. Mancarella. Computing assumption-based argumentation for multi-criteria decision making. Journal of Artificial Intelligence Research, January 2008. 14. J.F. Nash. Two-person cooperative games. Econometrica, 21:128–140, 1953. 15. M.J. Osborne and A. Rubinstein. Course in game theory. MIT Press, 1994. Cambridge, Massachusets. 16. I. Rahwan and L. Amgoud. An argumentation-based approach for practical reasoning. In Proc. AAMAS’06, pages 347–354. ACM Press, 2006. 17. I. Rahwan, S. Ramchurn, N. Jennings, P. McBurney, S. Parsons, and L. Sonenberg. Argumentation-based negotiation. The Knowledge

Engineering Review, 18(4):343 – 375, 2003. 18. M. Josefina Ramos. Industrial estates and regional development in selected Asian countries : a review of experience. United Nations Centre for Regional Development, 1991. 19. G. Richard Shell. Bargaining Negotiation Strategies for Reasonable People for Advantage. Penguin Books, 1999. 20. UNIDO. Industrial Estates Principles and Practices. Technical report, United Nations Industrial Development Organization, 1997. 21. J. van Veenen and H. Prakken. A protocol for arguing about rejections in negotiation. In Proc. ArgMAS’06, volume 4049 of LNAI, pages 138–153. Springer, 2006. 22. D. Zhang. Reasoning about bargaining situations. In Procs AAAI-07, pages 154– 159, 2007. 23. G. Zlotkin and J. S. Rosenschein. Negotiation and task sharing among autonomous agents in cooperative domains. In Proc. IJCAI, pages 912–917, 1989.

Attention! This is a preview.
Please click here if you would like to read this in our document viewer!